tech
April 2, 2026
AI labs enter compute wars
The more customers AI companies win, the more they spend. That's a problem no one's solved yet.
TL;DR
- Anthropic is experiencing success but is constrained by the high cost of compute power, leading to usage limits and outages for paying customers.
- AI labs must balance purchasing enough compute capacity to meet demand without overspending and damaging profit margins.
- CEO Dario Amodei acknowledges the difficulty of hedging against overbuying compute, preferring short-term customer loss over financial risk.
- Compute is essential for both serving customers and training new AI models, with labs scheduling training during off-peak hours to reduce costs.
- Despite falling compute costs due to chip and software efficiency, overall spending continues to climb due to skyrocketing usage (Jevons Paradox).
- Hyperscalers are expected to invest nearly $700 billion in AI capacity this year, yet this may not be enough to meet full demand.
- The AI race is shifting from model competition to a capital allocation problem, with the ultimate winners yet to be determined.
Continue reading the original article